JP2006012130A5

JP2006012130A5 -

Info

Publication number: JP2006012130A5
Application number: JP2005143944A
Authority: JP
Filing date: 2005-05-17
Publication date: 2011-07-28

Claims

A method of expressing a face image by a descriptor ,
Extracting a set of component images from the face image, each component image corresponding to a portion of the face;
Converting the pre-SL component image in the frequency domain,
Prior Symbol frequency domain, deriving a plurality of feature vectors using the transformed component images, and,
Processing the plurality of feature vectors to generate a descriptor representing the face image;
Including
The processing includes
An initial stage of discriminant transformation that combines and maps the derived feature vectors to generate a discriminated initial feature vector;
An intermediate stage of at least one discriminant transformation that combines and maps the determined initial feature vectors to generate a determined intermediate feature vector;
A final stage of discriminant transformation that combines and maps the discriminated intermediate feature vectors to generate a descriptor representing the face image;
Performing the three steps of discriminating conversion even without least including the including, Methods.

The method of claim 1, wherein at least one stage of discriminant transformation based on the feature vector comprises mapping the feature vector using dual LDA.

Extraction of the component images is based on the content of the image including the feature or area of a face of the face image, method towards the claim 1.

Transforming the component image into the frequency domain is based on Fourier transform,
Combining the feature vectors includes combining at least two of the Fourier coefficient real part, the Fourier coefficient imaginary part, and the Fourier coefficient amplitude of the component image, and / or the Fourier coefficient real part of different image components ; Fourier coefficient imaginary part, and includes coupling at least two of the amplitude of the Fourier coefficients, methods who claim 1.

The component image, the whole face image, 2 divided images該全body face image, divided into two images including, method towards 請 Motomeko 4 SL placing a portion image, and the portion image of the entire face image.

Deriving a feature vector of the Fourier coefficient imaginary part and the Fourier coefficients the real part of the entire face image and said portion image, and the characteristics of the amplitude of the Fourier coefficients of the first second divided image of the entire face image and the portion image deriving a vector, to derive an amplitude characteristic vector of the Fourier coefficients of the second 2 divided images of the entire face image and the portion image, including, methods who claim 5.

The initial stage of the discriminant transformation is: (i) the Fourier coefficient real part and the Fourier coefficient imaginary part of the whole face image, (ii) the amplitude of the Fourier coefficient of the two-part image of the whole face image, (iii) the Fourier coefficients real and the Fourier coefficients imaginary part of said portion image, and (iv) the combined amplitudes of the Fourier coefficients of the two split images of a portion image including the conversion, methods who claim 6, wherein .

Intermediate stages of discrimination transformations comprises mapping by combining the results of (v) (i) results and (ii) the results, and (vi) (iii) Results and (iv), according to claim 7 Law who described.

The final step of discriminating transformation, (vii) (v) results and combines the results of (vi) comprises mapping, methods who claim 8.

Before Kiutsushi image, for example, principal component analysis (PCA), linear discriminant analysis (LDA), independent component analysis (ICA), PCLDA or based on any of the dual LDA, any one of claims 1 to 9, 2. The method according to item 1 .

Conversion to the frequency domain comprises a Fourier transform, methods who claim 1 Symbol placement.

Deriving a feature vector of each component image comprises selecting specific predetermined Fourier coefficients, method towards the claim 11.

It comprises selecting a low horizontal / high vertical frequency components in at least the Fourier spectrum method towards the claim 12.

14. A method according to any one of the preceding claims, wherein the initial stage of the discriminant transformation further comprises normalization of the mapped derived feature vector.

15. A method according to any one of the preceding claims, wherein the intermediate stage of the at least one discriminant transformation further comprises normalization of the mapped initial feature vector.

The method according to any one of claims 1 to 15, wherein the final stage of the discriminant transformation further includes quantization of the mapped intermediate feature vector.

The derived expressions with how according to any one of claims 1 to 16, comprising comparing the the stored facial image representation, a method of performing facial recognition, detection or classification.

Representation which is the storage is derived using the way of any one of claims 1 to 16, a method of face recognition according to claim 17, detecting, or the classification performed.

Apparatus adapted perform method towards according to any one of claims 1 to 16.

Computer program for how according to any one of claims 1 to 16.

A computer-readable storage medium storing the computer program according to claim 20 .